aivrar

My Hermes Agent portable Windows application is getting a lot of attention. No need to use Docker or even have Python installed in your system, completely PORTABLE with a Native Windows GUI. github.com/aivrar/portabl…

@vllm_project vLLM is so awesome I made a Windows build for it. No need to use docker or have Linux/WSL. for RTX 30/40/50-series, pre-built wheel, Windows patchset, 10 KV-cache compression dtypes, OpenAI API server fixes, Rust frontend, and Rust tool parser support. github.com/aivrar/vllm-wi…

🚀 Qwen3.6-27B-NVFP4 is inference ready with vLLM on NVIDIA Blackwell GPUs. This checkpoint is optimized for Blackwell and reduces GPU memory requirements by ~2.5x for local AI with open-source models. 🧠 27B params, Hybrid Attention 📊 NVFP4 evals: 86.3 on MMLU Pro, 85.5 on GPQA Diamond 🛠️ Exclusively supported on vLLM as the runtime engine Get started from the Hugging Face checkpoint: huggingface.co/nvidia/Qwen3.6…

@YC1401 @FoxNews It's a great quote though, that's how I basically feel about the World.

@aivrar @FoxNews LOL, Believe no one... and not even no one!

BREAKING: Two people have climbed to the top of the Empire State Building in New York City, holding a banner from the skyscraper's antenna reading, "When the power of love beats the love of power, the world knows peace." As of now it's unclear how the pair reached the top of the building as police work to get them down from the spire, 1,454 feet above the ground.

@YC1401 @FoxNews Thanks! I didn't know that because on FB they just spam the Jimi Hendrix one.

@aivrar @FoxNews The original phrasing comes from 19th-century British Prime Minister William E. Gladstone: "We look forward to the time when the Power of Love will replace the Love of Power. Then will our world know the blessings of peace." Not Jimmy Hendrix lol

@MoundLore Wouldn't that be the dirt under our feet?

What’s the oldest thing you’ve ever touched?

@CinemaTweets1 The "Touched by an Angel" bit gets me every time.

One of the Funniest Performances in Cinema History

@coinbureau That's the kind of hype that makes people start looking to open source even more.

🚨ANTHROPIC CEO: OPEN SOURCE AI IS GETTING DANGEROUS Anthropic CEO Dario Amodei told lawmakers that open-source AI is moving down a “very dangerous path.” His warns that once powerful models are released openly, companies lose the ability to monitor misuse, revoke access, or update safety guardrails.

@hqmank Exact same for me, I was at 2% left and got the reset, happy Sunday indeed!.

I hit my Codex limit yesterday, but my weekly reset was coming soon. So I waited instead of pressing the reset button. Then OpenAI reset everyone anyway. I basically saved one extra reset by doing nothing.

@ThrillaRilla369 Biff.

I need a very specific tough sounding name for a tiny chihuahua puppy Not Rocky 🐕

@gofishh77 That couldn't work very efficiently without getting air into the bottom of that barrel.

Redneck ingenuity is always fun!

Playing with fire here, I know.

As a European, I am taking the climate pledge to NOT use air conditioning or other climate-destroying cooling devices Americans hate how strong we are (they are coddled and rely on artificial cool air) Who else is taking the pledge with me? Let's save the planet!

@Iberian_America I doubt they were random.

man this is like the 10th person who has been shot to death since I moved to tepito im starting to think tepito might not be safe idk

@Jackkk Not the worst life I guess.

Mark Zuckerberg reveals he's feeding his cows beer and macadamia nuts “On the ranch, one of my projects is I'm trying to create the highest quality beef in the world” “It's very low stakes, I’m not selling it but I'm very into the genetics of the cattle. We're trying to figure out how do you make it so that you basically can deliver the highest density diet to them” “We started growing macadamia trees because that kind of nut is extremely dense and they will eat a lot so they will put on weight and become fat quicker and become delicious” “The macadamia nuts have a lot of oil so you need to actually roast that. So now we need to design this whole process to roast the nuts so that way you can give them to the cows” “You want them to eat more. So then it's like how do you get them to eat more? Well it turns out alcohol is great for that because alcohol induces appetite” “That's actually why very high-end beef, they're fed beer. But okay, what's the right balance of beer versus water? I don't know. Let's let them choose. They get either as much cold beer as they want or as much room temperature water” “So now we're brewing all this beer and we're putting it out”

@DefiantLs Quinceañeras.

What's something that makes you think, "I'm too old for this shit"?

@lauriewired I'd totally game in the cloud, I need my GPUs for other AI stuff.

you’ll get mad at me for saying this…but cloud gaming is so obviously more economically efficient than physical hardware I think it’s going to be the default soon. your home console / pc is idle 90%+ of the day. meanwhile, data centers targets what, 5%, maybe at worst 10% idle. every second a cloud gamer isn’t gaming, that hardware is being used for someone else, training, etc. I think there should be a new measurement, something like cost-per effective FLOP hour that takes into account the TCO + effective utilization. If a gamer spends $500 on a GPU, uses it for 3 years, but it’s only fully active ~5% of that period…the cost-per relative FLOP hour is crazy high! Meanwhile, a $50,000 datacenter GPU might have a *LOWER* cost-per FLOP hour just because the effective utilization is 90+%.

@GithubProjects Then change it, what's the big deal?

I have a bad news

TurboQuant+ updates. 4.25→4.125 bpw, faster decode, lower KLD, crash fixed. NEED TESTERS: spent the last stretch hammering on my llama-cpp-turboquant fork and the numbers finally moved the way i wanted. this is the same-fork before vs after on qwen3.6 / 5090, so it isolates exactly what my fixes did: KLD vs f16: down 34-35%. the culprit was a mis-scaled centroid table (σ≈0.064 instead of exact Lloyd-Max). fixing it also cut PPL degradation from +0.55% to +0.16%. decode: +17% at short context, growing to +60% deep. a flash-attn launch-latency backport plus a fused-MMA decode path that reads the KV once per head-group instead of once per head. the 32K crash: fixed. turned out it was never an inference crash, just an int-to-size_t overflow in the perplexity tool. size: 4.25 down to 4.125 bpw, bit-identical quality. dropped a dead field that was just along for the ride. prefill: roughly flat. wasn't a target, it was already competitive. net: every quality, speed, and robustness axis moved the right way, and the block got smaller. loop's still live, chasing a clean -30% KLD bonus next. what i actually need now: testers. if you've got a 5090, or honestly any CUDA / HIP / Metal / Vulkan box, and you run local models, pull the branch and tell me what breaks. real hardware, real workloads, the messier the better. PR's here: github.com/TheTom/llama-c… Mixed results on pascal. need more info. plz let me know your results

@his4Everz Numbers aren't colors they're shades.

My fellow neurodivergents: What color is the number 9?

@morgan88877 Was? How about now?

Imagine how good the CIA's LSD was